Why the speech engine ships inside the app

The download is about 570 MB. For a dictation app that is enormous, and it is the first thing people ask about, so it is worth answering properly rather than in a footnote.

The size is the speech engine. It is in the download rather than on a server, and almost everything else about the product follows from that one choice.

The alternative, and what it costs

The normal way to build this is to record a buffer of audio, send it to a machine you operate, and paste back what comes out. It makes the app small, it works on any Mac, and it is why most dictation is priced monthly: somebody is running those machines whether you dictate or not.

It also sets three things you cannot then fix. The latency has a floor made of network, not computation. The app stops working when the connection does, which in practice means on a plane, on a train, on a client site that will not admit an unknown app, and on any ordinary bad day. And a copy of what you said exists, briefly, on hardware that is not yours.

What shipping it inside buys

A sentence is composed in about a fifth of a second, on the Mac, which is on the right side of the threshold where people keep talking rather than waiting to see what appears. It is the same speed with the router unplugged, because nothing was going anywhere. And there is no account, because there is nothing to have an account with.

The last one is the part that took the most work and is the easiest to state. Zumbo has no server for your voice. There is no account, no API key and no network request needed to transcribe. There is no data retention policy to read, because there is no database to retain anything in.

What it actually sends

Three things, and this is the complete list.

1. Your licence key, once, when you activate.
2. A version number, once a day, to check for an update.
3. A feedback message, only if you write one and press Send.

Block all three at the firewall and dictation is unaffected. That is a test you can run rather than a promise you have to accept, which is the point.

The cost of the decision

Being honest about the downsides: it needs Apple silicon and macOS 14 or later, because the model runs on the Neural Engine. On an Intel Mac this approach simply does not work, and Apple’s own dictation is the right answer there. The download is large enough to notice on a slow connection, once. And speaker labelling needs a second model, which is an optional add on of about 220 MB rather than something everybody downloads and most never use.

Those are real trade offs. They are the ones we would make again, because the alternative version of this app is one that stops working on a train and asks you to trust a policy about a recording of your own voice.

More on this in how Zumbo keeps your voice on your Mac and dictation with the Wi-Fi off.

• Speech engine inside the app
• 570 MB, downloaded once
• Works with the Wi-Fi off
• No account, no API key
• Unplug the internet and it works exactly the same.